Mid-term generation forecasting at DER level

From
Jump to: navigation, search

9 Mid-term generation forecasting at DER level (1 week ahead)

9.1 Rationale & Link to BEYOND Apps

The mid-term energy generation prediction at DER level is very important for forecasting and planning the maintenance of building assets. In traditional usage tied to distributed energy resources, mid-term forecasting is commonly used for regular maintenance scheduling, so the cost of interrupted operation is minimized. Shorter term forecasts are typically used in market participation. However, at the building level mid-term forecasts may be useful in e.g. regime change for installed equipment in the building. Effective mid-term generation forecasting is imperative for the establishment and continuous validation of energy management strategies and maintenance plans, at building level. By using mid-term forecasts of DER generation, the building manager can minimize the energy and comfort impact of scheduled maintenance. The mid-term generation forecasts will be available in the BEYOND AI Analytics Toolkit. The dedicated AI analytic with feed with valuable insights the Self-consumption optimization features of the Digital Twin environment (BEPO application), and the personal energy analytics PEASH application.

9.2 Overview of relevant implementations

As stochastic renewable resources, the DER production forecasts are largely impacted by the prime mover behavior – wind in case of wind power and insolation in the case of photovoltaics. Therefore, the meteorological forecasting is expected to play an important role. For the (very) short-term forecasts (below 6 hours of forecast horizon), the methods usually focus purely on statistical analysis of the production from the DER power plants. This is due to strong autocorrelation characteristics [1] [2]. Beyond 6 hours of forecasting horizon, using numerical weather forecasts as inputs to a DER forecasting system becomes beneficial. Essentially, predicting the DER output, at the time t, becomes a task of estimating an actual power at the time t+k, based on the forecasted value of input features such as wind speed or insolation.

P_(t+k)=f(x ̂_(t+k))

The process of predictor training relates to the configuration of the weights and other predictor inputs, so that the error measure between actual values and forecasted values is minimized. In effect, the well-trained predictor both learns the relation between the input feature (insolation or wind speed) as well as compensates for the NWP predictor errors (the difference between the forecasted values of the input features and the actual, observed values). From the above, it follows the mid-term energy generation forecasting can be viewed as a time series problem that is dealt with several approaches coming from classical statistics [2] (i.e. linear and non-linear models) and artificial intelligence (AI). The AI methods are able to learn from past behavior of variables and recognize the complexity and non-linearity in a certain dataset, as well as capable of including additional inputs. The statical approaches used for mid-term forecasting usually involve:

- the ARIMA (Autoregressive Integrated Moving Average) model, which originates from the autoregressive (AR) model, moving average (MA) model and

- the combination of AR and MA (ARMA). AR, MA and ARMA are applied to stationary series, while the ARIMA models are used for non-stationary series

- for seasonal time series, SARIMA models are used [3]

- stochastic models such as Markov chain [4]

The non-parametric AI based techniques used for load forecasting are usually artificial neural networks (ANN). As the tuning of predictor weights is an optimization process in nature, hybridization of ANN training process is often used so that Genetic Algorithms (GA) or Particle Swarm Optimization (PSO) is used to train the network instead of well-established algorithms such as the Back Propagation Algorithm (BPA) [5]. To perform the training and tune the predictor weights, the statistical forecasting approaches require large amounts of historical data to be present upfront as inputs to the process of supervised training [6]. When these data are not available, one can resort to modelled data such as the ones from meteorological reanalysis [7], or use an alternative forecasting method and learn online “as you go”. The time horizon for mid-term forecasting ranges from short-term to a limit of several weeks ahead. Moreover, as the time horizon increases, so do the forecast errors [8]. In literature, medium-term forecasting covers the time spectrum from several hours to 1 week ahead, and is used for unit commitment decisions, reserve requirement decisions and generator online and/or offline decisions. In the context of BEYOND, the forecast method will rely on meteorological forecasts from a NWP provider, and it has been decided to utilize a robust neural network method with NWP data as inputs.

9.3 Implementation in BEYOND

In BEYOND, the proposed methodology is based on artificial neural network as a predictor, that is trained on the input data in a supervised fashion, and then utilized as a trained predictor. The required data pre-processing steps are the following:

1) Filter and nullify outliers and erroneous entries (input cleaning)

2) Fill-in eventual missing values using interpolation through padding

3) Perform resampling to 1 item daily or hourly

4) Configure the season as a feature

5) Set the size of the acceptable error in the results

6) Configure additional synthetic feature as day of the week

7) Perform normalisation of input values if needed


The training step is an iterative procedure, where the input dataset is split into training and test sets. The training set is used to tune the predictor weights and the test set is used to evaluate the performance as it is not presented to the model during training. Typical size ratios range in 70 to 80%, 20 to 30% and 10 to 20% for training and test dataset sizes, respectively. Performing the forecast run on the test dataset provides a measure of the model performance, usually calculated as an accuracy The execution (usage) step of the analytics consists of running the input data through the trained predictor and obtaining the results.

9.3.1 Data inputs and Analytics Pipeline (incl. assumptions /limitations)

Given the expected forecast horizon, the numerical weather prediction data is required as an input, available at appropriate time ahead of the actual forecasted production realization. This input data is used for training and validating the model. In this development phase, we rely on the meteorological reanalysis data provided by MERRA-2 system [9] and available through the Renewables.ninja website [10] coupled with publicly available data from the ECMWF [11]. The PVGIS data, also publicly available, are used for cross-validation of the results. The analytics pipeline includes the required data pre-processing steps (see previous paragraph), training and testing of the model, as well as the evaluation of the results.

9.3.2 Analytics Libraries Employed

The key Python libraries used for data manipulation and data analytics are the following: - Pandas for time series management

- Numpy for numerical manipulation

- Sklearn for the implementation of the neural network

- Matplotlib for visualization of the training results

The first iteration of the prediction algorithm has been implemented as interactive Jupyter notebook.

References

[1] G. E. P. Box, G. M. Jenkins, and G. C. Reinsel, Time Series Analysis: Forecasting and Control, 4. Auflage. John Wiley & Sons, 2008.

[2] F. Petropoulos et al., “Forecasting: theory and practice,” ArXiv201203854 Stat, Oct. 2021, Accessed: Nov. 03, 2021. [Online]. Available: http://arxiv.org/abs/2012.03854

[3] V. G. Tran, V. Debusschere, and S. Bacha, “One week hourly electricity load forecasting using Neuro-Fuzzy and Seasonal ARIMA models,” IFAC Proc. Vol., vol. 45, no. 21, pp. 97–102, Jan. 2012, doi: 10.3182/20120902-4-FR-2032.00019.

[4] S. X. Chen, H. B. Gooi, and M. Q. Wang, “Solar radiation forecast based on fuzzy logic and neural networks,” Renew. Energy, vol. 60, pp. 195–201, Dec. 2013, doi: 10.1016/j.renene.2013.05.011.

[5] A. Baliyan, K. Gaurav, and S. K. Mishra, “A Review of Short Term Load Forecasting using Artificial Neural Network Models,” Procedia Comput. Sci., vol. 48, pp. 121–125, Jan. 2015, doi: 10.1016/j.procs.2015.04.160.

[6] G. Aburiyana and M. E. El-Hawary, “An overview of forecasting techniques for load, wind and solar powers,” in 2017 IEEE Electrical Power and Energy Conference (EPEC), Oct. 2017, pp. 1–7. doi: 10.1109/EPEC.2017.8286192.

[7] I. Staffell and S. Pfenninger, “Using bias-corrected reanalysis to simulate current and future wind power output,” Energy, vol. 114, pp. 1224–1239, Nov. 2016, doi: 10.1016/j.energy.2016.08.068.

[8] C. Monteiro et al., “Wind Power Forecasting: State-of-the-Art 2009,” Argonne Natl. Lab., Nov. 2009, doi: 10.2172/968212.

[9] “MERRA.” https://gmao.gsfc.nasa.gov/reanalysis/MERRA/docs/ (accessed May 30, 2018).

[10] “Renewables.ninja.” https://www.renewables.ninja/ (accessed Aug. 27, 2021).

[11] E. Andersson, “Medium-range forecasts,” ECMWF, Apr. 16, 2014. https://www.ecmwf.int/en/forecasts/documentation-and-support/medium-range-forecasts (accessed Mar. 25, 2022).


Back to BEYOND_Baseline_Analytics